Accuracy types explained

The field

Every result has two related fields: accuracy_type, a label, and accuracy, a number from 0 to 1. The label says how the point was placed; the number ranks results of the same type by how well the input matched.

The types

rooftop (1.0). The house number matched an address point from the National Address Database. The point is on the parcel or building. Safe for everything.

range_interpolation (0.8). No address point, but the street segment from TIGER/Line has a house-number range on the right side that contains the number, so the point is placed proportionally along that segment. Usually within a few tens of metres. Safe for mapping and district assignment; a rough guide for last-metre routing.

nearest_rooftop_match (0.7). The number is not in any range and not in the address points, but a nearby known number on the same street is. The point is that neighbour's. Fine for the neighbourhood, not for the door.

street_center (0.6). The street matched but no number could be placed. The point is the segment's midpoint.

place, zip, county, state (0.4 to 0.1). Only the containing area is known. The point is its centroid, which can be kilometres from the address. Never assign a district from these without a second look.

Using it

good = [r for r in results if r["accuracy_type"] in ("rooftop", "range_interpolation")]
review = [r for r in results if r["accuracy_type"] in ("place", "zip", "county", "state")]

Reverse geocoding adds distance_m, the distance from your point to the returned address, so you can set your own cut-off.

Why some states interpolate more

Address points depend on what each state has contributed to the National Address Database. Hawaii, Michigan, Mississippi, New Hampshire, Nevada and Puerto Rico currently have no state points and a few others are thin, so results there are mostly range_interpolation and say so. Street ranges cover all 50 states, DC and Puerto Rico.